openclaw - 💡(How to fix) Fix Bug: DeepSeek V4 thinking=high produces empty visible content, dispatches blank message to channel

Bug description

When DeepSeek V4 Pro (or V4 Flash) runs with thinkingLevel: "high" and the model finishes with stopReason: "stop", the API occasionally returns empty visible text while all output tokens went into the reasoning/thinking block. OpenClaw dispatches this empty message to the channel (Feishu, etc.) without detecting the edge case, resulting in the user seeing nothing on the channel side.

Root cause

DeepSeek V4 API in high thinking mode allocates all output tokens to the reasoning block (reasoning_content), and the content field in the final choice is empty or missing. OpenClaw converts this into a message with content: [{type: "thinking", text: ""}] and dispatches it — but the user-facing channel never displays thinking content, so the user sees a "lost" message.

The model DID complete successfully (1417 output tokens used for reasoning), but the visible answer was never produced.

Evidence

Session metadata

• OpenClaw version: 2026.5.7 (upgraded to 2026.5.18 in same session)
• Model: deepseek-art/deepseek-v4-pro
• thinkingLevel: high
• Session type: Feishu DM channel

Last assistant message (final turn, stopReason=stop)

{
  "stopReason": "stop",
  "content": [
    {
      "type": "thinking",
      "text": ""
    }
  ]
}

Content count: 1 (only empty thinking block, no text block)

Gateway dispatch log

21:22:48 [feishu] feishu[main]: dispatch complete (queuedFinal=true, replies=1)

The gateway dispatched the empty message to Feishu. The user received nothing visible.

Model API response (from trajectory)

{
  "ts": "2026-05-20T13:22:47.405Z",
  "type": "model.completed",
  "data": {
    "aborted": false,
    "timedOut": false,
    "usage": {
      "input": 35524,
      "output": 1417,
      "cacheRead": 1039872,
      "total": 1076813
    },
    "choices": []
  }
}

Note: choices array is empty, but output: 1417 tokens were consumed — all went to reasoning.

For comparison — previous turn (stopReason=toolUse, thinking works fine, text visible)

{
  "stopReason": "toolUse",
  "content": [
    {
      "type": "text",
      "text": "我看到了几个线索，让我逐个查一下。"
    },
    {
      "type": "tool_use",
      "name": "...",
      "input": {...}
    }
  ]
}

This proves the thinking→text conversion works fine for toolUse turns, only breaks for stop turns.

Proposed fix

When processing a stopReason: "stop" assistant message:

1. Detect empty visible content: If the message has thinking blocks but zero visible text, do NOT dispatch to the channel.
2. Auto-retry: Re-submit the conversation turn with a hint like "Please provide your final answer in visible text (not only in thinking)." OR re-submit with thinking disabled for this one turn.
3. Or fallback behavior: If thinking produced tokens but visible content is empty, try extracting/formatting the thinking content into visible text as a fallback (with a note that thinking was collapsed).

At minimum, the gateway should log a warning when dispatching a message with no visible text content.

Environment

• OpenClaw: 2026.5.7 → 2026.5.18
• Provider: deepseek-art via OpenAI-compatible API (https://api.deepseek.com)
• Model: deepseek-v4-pro (reasoning=true, thinking=high)
• Channel: Feishu (websocket)
• OS: Linux x64